Where the text you pasted goes
Three different things can happen to text you give an AI assistant. People mix them up, and each one is controlled by a different setting.
- Training. The conversation goes into a dataset used to train future versions of the model. There's no taking it back out: Anthropic says plainly that after you switch the setting off, your data stays in models already trained and in training runs already under way.
- Human review. Some chats are read by reviewers who rate the answers. Google's Gemini help pages advise against entering confidential data while activity saving is on, and keep reviewed chats for up to three years.
- Retention. Even with no training and no review, the conversation sits on the provider's servers for a while: to answer you, to prevent abuse, to investigate incidents.
The first is a toggle in the settings. On a personal plan, the second and third can't be fully switched off at all.
The data traffic light
The table below is our assessment by data type and by the conditions under which the assistant receives it. It rests on the help pages of OpenAI, Anthropic and Google and on two Ukrainian laws covered further down.
✓ — fine; ≈ — fine under the condition in the cell; ✗ — don't paste it. "Anonymised" means no names, company names, amounts or bank details that identify a party.
Why passwords are red even on a business plan: a password that has landed anywhere outside a password manager should be treated as compromised. It stays in the chat history, in data exports, on the screen of the colleague you showed the conversation to. The assistant never needs a password to answer.
A simpler rule for the team, if the table feels long: paste only what you'd be willing to forward to an unfamiliar contractor who hasn't signed an NDA. On a business plan that "contractor" has a contract with you; on a personal plan it doesn't.
Where to switch off training in three assistants
Menu labels as in the English interface. Sources: OpenAI "Data controls in ChatGPT", Anthropic "Is my data used for model training?", Google "Gemini Apps Privacy Hub".
A few details the table leaves out.
ChatGPT. The setting applies to the whole account on every device. If you use ChatGPT signed out, the choice is stored only in that browser. A temporary chat doesn't go into history or training, but if you save it, it becomes a regular chat and follows the account setting.
Claude. Since 28 August 2025 Anthropic has asked every Free, Pro and Max user to choose explicitly whether their chats can be used for training. Allow it and conversations are kept for five years; decline and it's 30 days. Incognito chats are never used for training, whatever the setting.
Gemini. Turning off Keep Activity stops both training and reviewer access. But, as Google puts it, chats are still used to answer you and to protect against abuse, including with human help. If activity is on, it auto-deletes after 18 months; you can change that to 3 or 36 months.
Three traps the menus don't show
1. The rating button. In all three services, a thumbs up or down is a separate permission that overrides training being off. OpenAI says that after such a rating the entire conversation may be used for training even if you opted out. Anthropic keeps a rated conversation for up to five years. Google takes the last 24 hours of your chats, and everything uploaded in them, along with the feedback. Tell the team plainly: we don't rate answers in work chats.
2. A deleted chat isn't always deleted. In Gemini, chats already seen by reviewers aren't deleted with your activity; they're kept for up to three years, disconnected from the account. With Anthropic, switching training off doesn't remove data from models already trained. So switch training off before the first work conversation, not after.
3. The setting is personal, not team-wide. On a personal plan each employee switches training off themselves, and you have no way of checking whether they did. Five people on five personal ChatGPT accounts means five separate settings and no control.
What Ukrainian law says
Personal data. The Law on Personal Data Protection requires a basis for any processing (Article 11): the person's consent, a contract with them, a legal duty, a legitimate interest. Pasting a customer list into an assistant is processing and needs a basis too. If customers agreed to their data being processed "to deliver the order", analysing it in a third-party AI service goes beyond that purpose.
Transfers abroad are a separate matter (Article 29). They're allowed to countries with adequate data protection: members of the European Economic Area, signatories of Council of Europe Convention 108 and countries on the Cabinet of Ministers' list. Otherwise, with the person's unambiguous consent, to perform a contract in their favour, or under the other conditions the article lists. For users in Ukraine, Gemini is provided by Google LLC, a US company; OpenAI and Anthropic are US companies as well. Before passing customer data to them, check what basis you rely on and whether your privacy policy mentions such a transfer.
Health, biometrics, ethnic origin, criminal records are special categories (Article 7). Processing them is prohibited apart from a closed list of exceptions, the first of which is the person's unambiguous consent. For a beauty salon, a clinic or an HR team this means a client's medical notes don't go into an assistant without that consent, on any plan.
Trade secrets. Under Article 505 of the Civil Code, information is a trade secret only if its owner took "measures adequate to the circumstances" to keep it secret. The law doesn't mention AI assistants. But in a dispute with a former employee or a competitor, you'll have to show what measures you took. An internal rule against pasting secrets into chats with training on, and a business plan, count in your favour. A habit of pasting everything into a free chat counts against you.
This section is an overview of the rules, not legal advice. If you process data on thousands of customers or special-category data, have a lawyer check your setup.
When you need a business plan
On the business plans of all three companies, training on your data is off by default:
- OpenAI doesn't train on data from ChatGPT Business, Enterprise, Edu or the API; for the API, the organisation owner can choose to turn data sharing on. Business is available from two users.
- Anthropic doesn't train on data from its commercial products, Claude for Work and the API. The owner of a team plan can disable the rating buttons for all members.
- Google doesn't use Gemini prompts, files and responses in Workspace to train models outside your domain without permission, and reviewers don't read them. The Gemini app is a core service in most Workspace editions.
For a small business the main difference isn't training as such: you can switch that off on a personal plan too. It's that the owner sets it once for everyone, and the data-processing terms are in a contract. You need a business plan if at least one of these is true:
- Two or more people use the assistant and you can't check everyone's settings.
- Contracts, financial documents or information you treat as a trade secret go into the assistant.
- You process customers' personal data and need a document describing how the provider handles it.
If none applies — you work alone and paste only copy and figures without names — a personal plan with training off is enough.
Our position
For a team of two or more, a business plan for an AI assistant is about control first and features second. Anyone can switch training off, but only an admin console can guarantee that everyone did and that nobody used the rating button. If contracts or customer data go into the assistant, the price gap between personal and business plans is what you pay to have something to show in a trade-secret dispute or a personal-data inspection.
The position doesn't hold if your assistant is a tool for website copy, emails without personal data and calculations on anonymised figures. Then everything you paste is already green in the table above, and a business plan protects nothing. How the models themselves handle business tasks is in our test of three AI assistants.
Checked on 7 October 2026 against OpenAI help pages (Data controls in ChatGPT, How your data is used to improve model performance), Anthropic (Privacy Center, the 28 August 2025 terms update), Google (Gemini Apps Privacy Hub, Generative AI in Google Workspace Privacy Hub), the ChatGPT pricing page, and the texts of Law No. 2297-VI and the Civil Code on zakon.rada.gov.ua.